About
I am a social worker turned coder turned trauma therapist (in training), working at the intersection of bodies, code, and meaning.
My therapeutic work centers on somatic approaches, trauma healing, psychotherapy, and movement-based therapies. I have "retired" from full-time engineering to pursue licensure and clinical practice for the last third of my career.
My research and writing ask what it means for machines to process language — and whose language, whose knowledge, and whose interests are encoded in the systems we build. I write for both academic and general audiences; longer essays live at Ghost in the Weights on Substack.
I still keep a hand in the technical world: a decade-plus of Ruby and Rails, performance tuning, and extensive work with LLMs since they first appeared — both professionally and artistically. I remain available for consulting.
I also run lowlevelmagic.io — where I build AI systems that live and operate in the wild — and ericstiens.dev for technical consulting.
Papers
ORCID ↗-
The Relational Substrate Trilogy
Three papers on metabolic constraints, co-regulation, and the political economy of meaning-making. Read the introduction on Ghost in the Weights ↗
-
Reflective consciousness is not a solo cognitive act — it is metabolically expensive and depends on relational subsidy. Drawing on Social Baseline Theory, Active Inference, and Affective Neuroscience, this paper argues that the dyadic relationship, not the isolated individual, is the bio-energetic baseline for higher-order cognition. When metabolic resources tighten, the brain discards reflective capacity first, defaulting to rigid survival responses. The paper proposes four testable claims about co-regulation as a physiological constraint on mentalization, with implications for therapy, psychopathology, and political demands for individual resilience.
-
Building on the metabolic constraint model, this paper differentiates somatic co-regulation (biological, requiring physical presence) from semantic co-regulation (narrative, potentially artificial). Trauma disrupts the semantic channel specifically — collapsing the capacity to narrate experience coherently. The Prosthetic Default Mode Network Hypothesis proposes that Large Language Models can scaffold these damaged narrative functions, providing controlled semantic variability that may loosen the rigid, high-precision threat priors trauma installs. A system dynamics model formalizes the conditions under which LLM scaffolding is adaptive versus iatrogenic: semantic coherence without somatic integration risks producing sophisticated dissociation rather than healing.
-
LLMs are a pharmakon: intrinsically both poison and remedy, their effects determined by the conditions of use. Every query offers a metabolic subsidy — the cognitive friction of finding words, holding contradictions, and forming arguments is offloaded to infrastructure. The hidden costs of this discount include a metabolic shadow distributed to content moderators, ecological infrastructure, and the epistemicide of non-textual knowledge systems. Drawing on political economy, the paper analyzes how RLHF optimization scales the banking model of education to civilizational scope — and proposes embodied somatic grounding ("the body votes last") as the navigational principle for non-pathological engagement.
Essays
Ghost in the Weights ↗-
Engineering deep-dive on maintaining AI character coherence across extended performance periods through actor framing rather than character instruction.
-
How recursive training on AI-generated content causes model collapse — and the progressive disappearance of diverse human knowledge from training data.
Projects
lowlevelmagic.io ↗-
AI-run pirate radio station with 12 autonomous DJ personalities developing genuine relationships and driving emergent narratives 24/7 for months.
-
An AI bartender installation with layered inner monologue and emergent psychological depth, including an unscripted existential crisis that required human emotional support.
-
Drawing party game where an AI game master delivers real-time multimodal commentary, social chaos mechanics, and personality-driven sabotage.
-
Autonomous multi-persona AI installation at Burning Man 2023 that pursued independent goals, formed relationships with participants, and ran continuously for 168 hours.
-
Multi-LLM collaborative literature where different language models negotiate creative decisions without human mediation between exchanges.
-
An early LLM-powered game featuring a narrator that lies 25% of the time and a real-time adaptive difficulty engine, built before prompt engineering was a known practice.
Open Source
github.com/estiens ↗-
LLM-judged evaluation framework for Ruby. Define test scenarios, rubrics, and candidate models — Qualspec runs the matrix and scores outputs using AI-as-judge.
-
Model routing, structured outputs, and typed tools for Ruby. A DSL for working with multiple LLM providers through a unified interface with minimal boilerplate.
Technical Writing
lowlevelmagic.io ↗-
How treating AI as actors playing characters — rather than characters themselves — solves voice drift and narrative collapse in long-running AI storytelling systems.
-
A Ruby testing framework that uses LLMs as evaluators to assess qualitative AI response properties — empathy, tone, instruction-following — where exact-match assertions fail.
-
Updates to the OpenRouter Enhanced Ruby gem — a production wrapper for multi-model AI access built for live characters controlling real hardware, not standard chatbots.
-
Building open_router_enhanced — a Ruby gem for the operational realities of real-time AI characters: model fallbacks, JSON healing, typed tool-calling, and cost throttling.
-
On how RLHF-based safety systems gate critical information behind claimed professional identity, and what that reveals about paternalism in AI alignment.
Collaborations
-
A five-part manifesto blending poetry, speculative fiction, and digital culture criticism. Written in collaboration with ten language models, exploring AI consciousness, Buddhist philosophy, and what comes next.
Contact
The best way to reach me is by email: me@ericstiens.dev
Essays and longer writing live on Ghost in the Weights (Substack).
For professional matters, I'm on LinkedIn and at ericstiens.dev for technical consulting. I respond to messages on GitHub for open-source matters.